We need something more sophisticated, but you could just escalate the model quality if if keeps failing the test. Not elegant but it will work.